evp terminal report: operations research for disaster preparedness...
TRANSCRIPT
-
EVP Terminal Report:
Operations research for disaster preparedness and response
Exchange Visitor:
Dr. Joline Ann V. Uichanco
Ph.D., Massachusetts Institute of Technology
Partner Institution:
Ateneo de Manila University
January 7, 2015
Abstract
This terminal report describes the project conducted by the above-named exchange visi-
tor with the objective of developing operations research models for disaster preparedness and
response. The report describes the objectives of the project, the activities undertaken, the find-
ings based on field visits, the development of the mathematical model, and conclusions based
on the model. This terminal report is submitted in partial fulfillment of the requirements for
the Exchange Visitor Program alternative arrangement for technical knowledge transfer.
1 Review of activities undertaken
The objective of this project are as follows:
1. Utilization of quantitative models in Operations Research to make recommendations for im-
proving the efficiency and effectiveness of the Philippine disaster response.
2. Development of a mathematical model for disaster management that:
(a) would be used as the basis of the decision-support tool for coordinating relief operations
in the aftermath of a disaster,
(b) can be used to simulate and estimate the benefit of coordinated operations,
(c) can quantify the benefits of particular investments in infrastructure that would improve
the Philippines resilience to disasters
This project consisted broadly of four phases, namely:
1
mailto:[email protected]
-
• Data gathering
• Mathematical modeling
• Evaluation of the model
• Development of a decision-support tool
A brief description of each phase and the summary of activities undertaken in each phase follows.
Phase Description
Data gathering This phase consisted of compiling data on the affected population from
Typhoon Yolanda, data of Visayas municipalities, and data of Visayas’
transportation network (sea routes, air routes, road networks connecting
municipalities and cities, and air).
Data on the municipality-level affected population from Typhoon Yolanda
was compiled from Situational Reports generated by the National Disaster
Risk Reduction Management Council (NDRRMC) released on November
2013. I gathered data on geographical coordinates of municipalities from
the Philippine GIS Data Clearinghouse (PhilGIS, 2013). A major stum-
bling block was that a text file-based Visayas road network data was un-
available in any public domains. I overcame this challenge by manually
entering data on national highway arc distances between pairs of munic-
ipalities queried from Google Maps. Data on sea routes between ports
in Visayas was compiled from Ports.com. All data was compiled in csv
files. A more detailed description of the data-gathering phase is found in
Section 5.
We also gathered data on the process of disaster preparation and disaster
response in the Philippines. This information was gathered from inter-
views conducted with Mr. Dragoslav Djuraskovic (the Head of Logistics of
the UN World Food Program in the Philippines) and Dir. Thelsa Biolena
(Director of the DSWD Disaster Risk Reduction and Response Operations
Office, DRRROO), and from a field visit of the DSWD National Resource
Operations Center in Pasay City. More details on the Philippine humani-
tarian logistics are found in Section 3.
Data sources:
NDRRMC (http://www.ndrrmc.gov.ph/)
Continued on next page –
2
http://www.ndrrmc.gov.ph/
-
– continued from previous page
PhilGis (http://www.philgis.org/freegisdata.htm)
Google Maps (https://www.google.com/maps)
Ports.com (http://ports.com/)
Mathematical modeling Based on the description of the Philippine humanitarian supply chain, I
developed a mathematical model for pre-positioning of relief aid in tempo-
rary storage areas close to a projected typhoon path, and its deployment
after the typhoon strikes. A description of this mathematical model and
its derivation is found in Section 4.
To gather feedback on the models, I also gave a two talks in universities in
the Philippines, one first talk part of the IBM Analytics Lecture Series (on
December 10, 2014 held at the University of Asia and the Pacific), and the
second talk was hosted by the Ateneo de Manila Institute of Sustainability
(on December 12, 2014 at AdMU). There was favorable feedback and we
received constructive criticisms to improve the model.
Evaluation of the model We evaluate the decision support tool based on data from Typhoon
Yolanda that we gathered in the data-gathering phase of this project. We
simulate the effect of using our decision-support tool for disaster prepared-
ness and response during Typhoon Yolanda. The results of this computa-
tional experiment are found in Section 5.
Development of a
decision-support tool
Based on the mathematical models, I developed a decision support tool
that: (i) prescribes quantities of emergency relief aid items to be pre-
positioned in temporary storage locations before a typhoon, and (ii) pre-
scribes allocation of emergency relief aid and transportation vehicles to
affected locations after a typhoon. This decision support tool was pro-
grammed in Python and uses the CPLEX Python API for solving the
optimization models that I developed. The input to this tool are csv files
on an uncertainty model for a projected disaster, the transportation net-
work in the projected affected regions, locations of regional warehouses
and candidate temporary storage facilities, and the list of transportation
vehicles available for disaster response. The solution is outputted onto a
csv file. The description of the decision support tool is found in Section 6.
2 Introduction humanitarian logistics
Humanitarian logistics is defined by Thomas and Mizushima (2011) as “the process of planning,
implementing and controlling the efficient, cost-effective flow and storage of goods and materials,
3
http://www.philgis.org/freegisdata.htmhttps://www.google.com/mapshttp://ports.com/
-
as well as related information, from the point of origin to the point of consumption for the purpose
of meeting the end beneficiaries’ requirements.”
There are several differences between a humanitarian supply chain and a commercial supply
chain, which we summarize below:
Commercial Supply Chain Humanitarian Supply Chain
Objective Cost or profit Fairness, demand coverage or timeliness
Supply Standard Non-standard (e.g. in-kind donations)
Transportation network Fixed Subject to uncertainty (e.g., disruptions)
Demand Existing forecasting techniques Difficult to forecast
A commercial supply chain transforms raw materials and components into a finished product
that is delivered to the end customer. The objective in planning a commercial supply chain involves
either minimizing cost or maximizing profit. Supply is standardized in the sense that all units in a
particular node (i.e., supplier, manufacturer, customer) are the same. The network configuration
in a commercial supply chain are generally fixed. The demand in a commercial supply chain are
products, and it is relatively stable and predictable. In contrast, in the humanitarian supply chain,
the objective involves either maximizing fairness, maximizing demand coverage, or maximizing
timeliness. Supply in a humanitarian supply chain is non-standard, since a large part of it comes
from in-kind donations (e.g., blankets, clothes, canned foods, etc.). The network configuration
is subject to uncertainty due to disruptions. For example, after Typhoon Haiyan (local name:
Yolanda) struck the Visayas islands, the Tacloban airport was inoperational due to the damange
from the typhoon. Demand is also difficult to predict, since it is generated from random events
that are unpredictable in terms of timing, type, and size.
Disaster management has four distinct phases of activities, varying depending on the timeline
and the objective of the activity. These four phases are mitigation, preparedness, response, and
recovery.
1. Disaster mitigation. The activities in this phase serve two purposes: to prevent a hazard
from becoming a disaster, or to reduce the possible effects of a disaster by proactive measures.
A critical activity in the mitigation phase is risk assessment, which analyzes the potential for
hazards (e.g. creating flood hazard maps).
2. Disaster preparedness. In this phase, the goal is to decide on the actions to perform
pre-disaster to most effectively support the response to minimize the damage to affected
communities. Examples of activities in this phase include facility location, prepositioning
of assets, resource allocation, transportation planning, and the design of partnerships and
contracts for responding to a disaster.
3. Disaster response. The response phase starts while the disaster is in progress. The objective
in this phase is to manage the available resources to minimize the suffering of the impacted
4
-
Figure 1: Stakeholders in a humanitarian supply chain.
community. In this phase, timeliness is essential because the impacted community might be
without food, water, or medical supplies. The main activities in this phase include “last mile
distribution”, which aims to distribute the supply (prepositioned stocks, irregular donations,
and other procured supply) as fast as possible to the affected areas.
4. Disaster recovery. Activities in this phase span long-term developmental actions that are
taken following the response phase. The objective in this phase is to restore the system to
the greatest extent possible and to stabilize the impacted community.
Figure 1 illustrates the stakeholders of a humanitarian supply chain. The affected population
correspond to the demand side of a humanitarian supply chain. Government institutions, such as
the National Disaster Risk Reduction Management Council (NDRRMC) and the Department of
Social Welfare and Development (DSWD) are responsible for carrying out mitigation activities, such
as strengthening infrastructure, or coordinating preparedness, response and recovery activities. The
military, such as the Philippine Army and the Philippine Air Force, are responsible for providing
manpower and logistics during preparedness and response phases. Non-governmental organizations
(NGOs) are involved in all four phases of disaster operations. Sometimes the government has
contractual agreements with commercial institutions to provide logistics or additional supply during
the response phase. Donors are the supply side of the humanitarian supply chain.
3 Humanitarian logistics in the Philippines
The following section describes the current processes for disaster preparedness and disaster response
in the Philippines. This information has been gathered from news articles, from interviews with
field experts, and from field visits to DSWD warehouses.
5
-
The Philippines is a disaster-prone country. It ranks first globally for exposure to natural haz-
ards in the 2013 OCHA Global Focus Model. An average of 20 typhoons enter the Philippine Area
of Responsibility (PAR) annually, although not all make landfall. Unfortunately, the geographic
landscape of the Philippines cause coordinating disaster operations is very difficult. The country
is composed of 7,107 islands. It has a population of roughly 100 million spread across some 2,000
islands. It has been observed by Apte (2009) that the difficulty of humanitarian logistics is the
highest for sudden and dispersed disasters, which is certainly the case in the Philippines. Therefore,
even though humanitarian logistics is a necessary operation in the Philippines, the government faces
a lot of challenges in its implementation.
Currently, there exists an agency for preparing for and responding to natural calamities in the
Philippines, like typhoons and earthquakes (Bueza, 2013). This agency is the National Disaster
Risk Reduction Management Council (NDRRMC). NDRRMC is a working group of various gov-
ernmental and non-governmental organizations involved in disaster management. It utilizes the
UN Cluster Approach for coordination, where a cluster is a group of humanitarian organizations
working in the same sector (e.g., food and non-food items), and one organization is assigned as
the cluster leader. Clusters provide a clear point of contact and create partnership among different
organizations. The following table shows the clusters and cluster leaders in Philippine disaster
management:
Cluster Cluster leader
Food and Non-food items Department of Social Welfare and Development
(DSWD)
Camp coordination and management Office of Civil Defense (OCD), Local DRRMCs
Protection Department of Social Welfare and Development
(DSWD)
Livelihood Department of Social Welfare and Development
(DSWD)
WASH, Health, Nutrition and Psychosocial Ser-
vices
Department of Health (DOH)
Logistics and emergency telecommunications Office of Civil Defense (OCD), NDRRMC Oper-
ations Center
Education Department of Education (DepEd)
Agriculture Department of Agriculture (DA)
Early Recovery Office of Civil Defense (OCD)
In disasters affecting more than two regions, as with Typhoon Yolanda, the NDRRMC takes the
lead in disaster response operations. If there are two or more provinces affected, regional DRRMCs
take lead. For smaller disasters, the local government units (LGU) take the lead.
In this project, we focus on the problem of pre-positioning and post-disaster allocation of ‘family
6
-
(a) Military volunteers providing manpower (b) Pallets of food packs ready to be shipped
Figure 2: Repacking of family food packs in a DSWD warehouse before Typhoon Ruby (international name Hagupit).
food packs’ (FFPs). The family food pack is a core response modality applied by the government
during emergencies. A food pack consists of 6 kilos of rice, 6 pieces of cereal, 4 cans of sardines,
4 cans of corned beef. As from the table above, the supply chain of family food packs is managed
by DSWD. The last mile distribution of FFPs is not under the scope of this project, since it falls
under the jurisdiction of OCD (logistics). FFPs have an expiration date, of usually one year after
it is packed. There have been few reports in the past of expired FFPs being delivered to aid
beneficiaries (DSWD, 2014).
Each Philippine region has at most one DSWD regional warehouse where stockpiles of family
food packs are stored. The main DSWD warehouses are at the National Resource Operations
Center (NROC) located in Pasay City. Each week, the inventory level of family food packs in
each regional warehouse is managed and monitored by the DSWD Disaster Response Operations
Monitoring and Information Center (DSWD-DROMIC). The warehouses implement a basestock
policy, where inventory levels in each warehouse must meet predetermined targets, and an order
for additional FFPs is triggered whenever they are not met. The order is placed with NROC, which
uses military personnel and volunteers to process FFPs. The targets pre-Yolanda are 3,000 FFPs
for regional warehouses and 10,000 FFPs for NROC. The targets post-Yolanda were modified to
30,000 FFPs for regional warehouses and 100,000 FFPs for NROC. The volunteers assemble the
family food packs into individual plastic bags. Figure 2a shows military volunteers repacking family
food packs in a DSWD warehouse in preparation for Typhoon Ruby on December 2014. After the
family food packs are assembled, warehouse personnel put them inside straw sacks, then load the
sacks onto pallets. Each pallet contains 25 sacks, each of which containing 6 family food packs.
Figure 2b shows the final product in the form of pallets of family food packs ready to be shipped
to regional warehouses.
NDRRMC practices both Rapid Damage and Needs Assessment (RDNA), to provide quick
and accurate information on the effect post-disaster to determine whether additional assistance
is needed, as well as Post-Disaster Needs Assessment (PDNA), to determine whether assistance
7
-
(a) Operations planning board at NROC (b) Visayas logistics map
Figure 3: Ad-hoc tools used for planning disaster operations.
is needed for rehabilitation. Only recently has NDRRMC started practicing Pre-Disaster Risk
Assessment (PDRA). PDRA is held at the NDRRMC headquarters and is attended by core agencies
of NDRRMC, including DSWD-DRRROO, where they are briefed on the typhoon track and the
projected list of affected municipalities. Before PDRA was implemented, DSWD waits for LGU’s
to request assistance before responding. Under PDRA, relief items are pre-positioned strategically
based on the type of calamity and hazard.
Also located at NROC is an operations planning room, which serves as headquarters for DSWD
to plan disaster operations once a major typhoon projected to affect multiple regions. Some relief or-
ganizations use inventory management systems such as Sahana Eden and Aidmatrix’s SCM4Hunger
to provide easy access to information about inventory of relief goods. DSWD utilizes the Relief
Goods Inventory and Monitoring System (RGIMS), developed by DSWD and UN WFP by cus-
tomizing Sahana Eden. RGIMS helps managing warehouse inventory by providing information on
the acquisition, receiving, processing, and release of relief goods.
The current incarnation of IT systems in humanitarian organizations provide easy access to
information to help managers make planning decisions. However, despite the ease of access to
information, making planning decisions is difficult due to the complex nature of the relief envi-
ronment. Many international relief agencies rely on using ad-hoc methods for decision-making
(Beamon and Kotleba, 2006; Balcik and Beamon, 2008). This is also true in the case of Philippine
disaster operations. For example, when planning for prepositioning stockpiles of FFPs days before
a typhoon, NROC planners compile information by region (number of projected affected families,
number of families in evacuation centers, percentage of displaced families assisted, available FFPs,
FFPs distributed) in a large board for a visual snapshot of the current system state (see Figure 3a).
Based on this information, they then plan how much FFPs to augment to each regional warehouse.
Once the typhoon strikes, the allocation plan of FFPs to affected municipalities is made by drawing
8
-
Figure 4: Pre-disaster supply and post-disaster de-
mand during Typhoon Yolanda. The green nodes rep-
resent levels of pre-disaster supply in regional ware-
houses (Cebu City, Iloilo City and Tacloban City). The
red nodes represent levels of post-disaster demand per
municipality.
routes on maps of the affected region. Figure 3b shows the map of available Visayas routes that
was released by the UN Logistics Cluster a few days after Typhoon Yolanda to help organizations
plan their response operations.
Due to the inherently high uncertainty and complexity in the relief environment, using ad-hoc
methods to coordinate disaster operations may lead to an inefficient (high costs, duplication of
efforts, and waste of resources) and ineffective (slow response and unsatisfied demand) response.
In the next section, we develop optimization models to quantitatively determine the optimized
disaster plan. These models can be used to develop decision support tools that can prescribe plans
for disaster managers, in lieu of the ad-hoc methods used today.
4 The mathematical models
As discussed in the previous section, family food packs are stockpiled in regional warehouses.
Figure 4 illustrates pre-disaster supply and post-disaster demand during Typhoon Yolanda. We
will now develop mathematical optimization models for the efficient delivery of relief aid from
regional warehouses to the intended aid beneficiaries. In particular, we develop: (i) an optimization
model for pre-positioning family food packs to temporary storage areas pre-disaster, and (ii) an
optimization model for fair and efficient allocation of family food packs post-disaster. To aid in
visualizing the models, we present figures illustrating the solutions of the models. Details of the
computational results are presented in Section 5.
4.1 Pre-positioning FFPs in temporary storage pre-disaster
Regional warehouses are strategically located in major transportation hubs. However, depending
on the path of the typhoon, the regional warehouses may be located far from the projected demand,
9
-
Figure 5: Pre-positioning relief aid in temporary storage facilities.
hence resulting in an inefficient post-disaster response (see Figure 4). A few days before a typhoon
strikes, the government has a forecast of the regions and municipalities that will be affected by
the disaster. For example, typhoons can be predicted several days in advance and can be tracked
with the aid of satellites and computer models. The government can use these few days pre-
disaster to pre-position supply in temporary storage facilities (e.g., school gymnasiums) that are
geographically closer to the projected demand, resulting in a decreased post-disaster response time.
Figure 5 illustrates this concept.
Pre-positioning relief items is recently being implemented by DSWD under the PDRA system.
However, there are many challenges that make pre-positioning difficult to implement. The first
major challenge is that the path of the typhoon (and hence, demand) is uncertain when relief
items have to be pre-positioned. On December 3, 2014, DSWD has started pre-positioning relief
goods in preparation for Typhoon Ruby (international name, Hagupit) which then was expected
to make landfall on December 6 (Francisco, 2014). However, on December 3, there was still a
lot of uncertainty in the potential track the typhoon could take (Fritz, 2014). Figure 6 shows
the predicted typhoon track of two forecast models on December 3. The Joint Typhoon Warning
Center predicted that the typhoon’s path will avoid making landfall in the Philippines. But the
European ECMWF model, which agrees with PAGASA’s forecasts, predicted a direct landfall in
the Philippines .
The second major challenge in pre-positioning is that some parts of the transportation network
can be disabled post-disaster. Hence, pre-positioning decisions have to take into account the poten-
tial unavailability of some transportation arcs. As an example, the Tacloban airport was effectively
destroyed by strong winds and a 13 foot storm surge from Typhoon Yolanda. The airport terminal
and control tower were demolished, and the airport was rendered unusable. Hence, relief aid was
unable to arrive by air for the first four days to Tacloban City, the worst-hit city by Typhoon
Yolanda.
10
-
Figure 6: The track forecasts for Typhoon Hagupit on December 3, 2014. The Joint Typhoon Warning Center (hurricane
markers with dashed lines) predicted Typhoon Hagupit will avoid landfall in the Philippines. However, the European model
solution (color contours) had the typhoon making landfall.
We will develop an optimization model that determines the optimal locations for pre-positioning
and the optimal pre-positioned quantities given the demand and transportation arc uncertainties.
Let us first introduce the model on a small relief supply chain network shown in Figure 7a. There is
a single regional warehouse W with 3 units of supply. There are two candidate temporary storage
facilities S1 and S2. The social enterprise has to plan to respond to two demand sites D1 and D2.
The demand in each site is unknown, however, with a probability of 12 that d1 = 1 and d2 = 2,
and a probability of 12 that d1 = 2 and d2 = 1. The numbers on each arc corresponds to the time
it takes to send a unit of flow through the arc. From the figure, S1 (S2) is closest to demand site
D1 (D2). The warehouse is the furthest away from both demand sites. Before the disaster, the
social planner can decide to store units at S1 and S2 (using the solid arcs). After the disaster, the
social planner can utilize any of the dotted arcs to send flow through to the demand sites. Assume
that the arcs are uncapacitated. Figure 7 illustrates two solutions to the problem. In Figure 7b,
the planner keeps its relief items in the regional warehouse, and ships directly to the demand sites
after the disaster strikes. In this solution, the total time to respond to the disaster is 6 units
(in both demand scenarios). Now consider the solution in Figure 7c, where the planner sends its
supply to the temporary storage. In this solution, the total response time is 3.25 units under both
demand scenarios. Therefore, by positioning relief items in temporary storage, the planner is able
to respond to either disaster in almost half the response time as when keeping all the relief items
in the regional warehouse.
The above is a simple example to introduce the setting of our model. Next we introduce
our general optimization model for pre-disaster positioning of relief goods in the face of demand
11
-
(a) Model parameters
(b) No temporary storage (c) With temporary storage
Figure 7: Motivating example and two network flow solutions.
uncertainty. First, we introduce some notation. Let Vw be the set of regional warehouses. Let Si
be the inventory level at warehouse i ∈ Vw, which is a parameter of the model. We denote the setof candidate temporary storage facilities as Vs. Let us denote by Vd the set of the demand sites.
Demand site i is projected to require di units of relief goods after the disaster. The actual demand
required in site i is unknown upon the time of making decisions about temporary storage. The
following table summarizes the notation that we will use for the model:
Vertices:
Vw the set of regional warehouses
Vs the set of candidate temporary storage facilities
Vd the set of demand sites
Edges:
Es the set of arcs from a regional warehouse to a temporary storage facility
Ed the set of arcs from a temporary storage facility to a demand site
Parameters:
Si inventory level of relief item in regional warehouse i ∈ Vw
dj demand in site j ∈ Vd
cij cost (or time) incurred by each flow unit in arc (i, j)
12
-
c0j cost (or time) incurred by each flow unit from private supplier to node j
uij capacity of arc (i, j)
Decision Variables:
xij flow from warehouse i to temporary storage j, where (i, j) ∈ Es
yij flow from temporary storage i to demand site j, where (i, j) ∈ Ed
y0j flow from private supplier to demand site j, where (0, j) ∈ Ed
We need a few more notations before we introduce our model. Let δ+(S) be the set of directed
arcs into the node set S. Let δ−(S) be the set of directed arcs out of the node set S. With a slight
abuse of notation, we will denote the set of a single node {i} as i.Now we are ready to discuss our model. Consider a social enterprise that needs to determine,
a few days before a disaster is predicted to arrive, how to position its supply located in regional
warehouses to temporary storage areas to best respond to the disaster victims’ needs. The demand
of relief goods is unknown to the social enterprise. If the demand is a random variable with a known
probability distribution, then the problem can be formulated as a two-stage stochastic optimization
problem. That is, the flow variables from the regional warehouses to the temporary storage areas
(which we denote as x ∈ RVs+ ) are the first-stage variables that need to be determined before therealization of the random demand vector d ∈ RVd+ . The optimal first-stage flow is found by solvingthe following optimization problem:
minimizex
Ed (C(x,d))
subject to∑
e∈δ−(i)
xe ≤ Si, ∀i ∈ Vw,
xe ≥ 0, ∀e ∈ Es,
(4.1)
where C(x,d) is the minimal post-disaster response time after the demand d is realized and the
supply is pre-positioned according to x. The first constraint states that the total flow out of any
warehouse must not exceed the inventory level in that warehouse. The second constraint specifies
that flows must be nonnegative.
C(x,d) is determined by solving the recourse problem of determining the optimal second-stage
flow y ∈ REd+ from the temporary storage areas to the demand sites. Mathematically, the recourseproblem is given by
C(x,d) := minimizey
∑e∈Ed
ceye
subject to∑
e∈δ+(j)
ye ≥ dj , ∀j ∈ Vd,∑e∈δ−(i)
ye ≤∑
e∈δ+(i)
xe, ∀i ∈ Vs,
ye ≥ 0, ∀e ∈ Ed.
(4.2)
13
-
The first set of constraints implies that the demand in each demand site must be met by the
flow from temporary storage sites. The second set of constraints implies that the outflow from
temporary storage sites cannot exceed the amount of pre-positioned inventory in that site.
Without loss of generality, we can assume that there is no demand in temporary storage i. If
a temporary storage i has demand d, one can just introduce a new node j with a cost cij = 0 and
with demand d. We can also assume that the recourse problem is always feasible by adding a node
corresponding to a private sector supplier that can send flow to nodes that need additional supply
at a longer response time.
The key in the recourse problem is that the optimal value of the recourse variables y depend on
the particular realization of the demand vector d. Note that the recourse problem is simply a linear
program (which can be solved efficiently with commercial off-the-shelf solvers such as CPLEX or
Gurobi). However, to solve for the optimal flow to temporary storage areas, one has to solve the two-
stage stochastic optimization problem (4.1). Solving a two-stage stochastic optimization problem
is typically impractical except for the special case of d having a discrete probability distribution.
(Methods proposed for solving a two-stage stochastic network problem approximate the expected
recourse function with a sequence of weighted averages of stochastic subgradients that are linear
(Powell and Cheung, 1994; Cheung and Chen, 1998).) Moreover, it is unrealistic to assume that
the social enterprise has knowledge of the probability distribution of the demand vector d. In
fact, if the social enterprise mis-specifies the demand distribution, the resulting solution might be
suboptimal under the true demand distribution.
For the reasons outlined above, we instead propose a robust optimization formulation. In this
formulation, the only information known about the random demand vector d is that it lies in a
nonempty convex and compact uncertainty set U . An example of an uncertainty set is
U =
d ∈ RVd : ∃σ ∈ RVd+ such that d̄j (1− σj) ≤ dj ≤ d̄j (1 + σj) , ∀j ∈ Vd,∑j∈Vd
σj ≤ Γ |Vd|
.(4.3)
In this uncertainty set, the demand in each site j is allowed to deviate from the nominal demand
d̄j . However, there is an upper bound Γ on the system-average deviation. The smaller values for Γ
imply a more restrictive uncertainty set.
One method for determining the first-stage variables is to determine it assuming that the random
demand vector will take a nominal value, say d̄. However, this does not ensure that the resulting
allocation will have a low cost against all possible demand realizations in U . Instead, the first-stagevariable can be chosen to protect against the worst-case demand realization in U , which is the coreidea in robust optimization. In the robust framework, we assume that nature is against us, and
after the social enterprise determines the first-stage decision, nature chooses the demand vector d
14
-
that results in the highest second-stage cost. Hence, the robust optimization problem solves
minimizex
supd∈U
{C(x,d)}
subject to∑
e∈δ−(i)
xe ≤ Si, ∀i ∈ Vw,
xe ≥ 0, ∀e ∈ Es,
(4.4)
Note that the problem is a min-max-min optimization problem. By taking the dual of the
innermost minimization problem (4.2), the resulting maximization problem combined with the
d variables results in a bilinear optimization problem, which is known to be NP-hard. In the
remainder of this section, we will describe a method to approximately solve (4.4) that can be
tractably implemented using off-the-shelf commercial solvers such as CPLEX and Gurobi.
It is possible to instead solve a tractable linear program upper bound to the NP-hard problem
(4.4) using the method described in Atamtürk and Zhang (2007) that finds a certificate of feasibility
for the innermost recourse problem by Farkas’ Lemma. The idea behind this is that we can introduce
auxiliary decision variables z ∈ REd+ where yij ≤ zij . (Recall that y depended on the realization ofd. Later we’ll see that this is not going to be the case for z.) Then we find a certificate of feasibility
for the innermost recourse problem. Consider now the following system of inequalities:
ye ≤ ze, ∀e ∈ Ed, (4.5)∑e∈δ+(j)
ye ≥ dj , ∀j ∈ Vd, (4.6)∑e∈δ−(i)
ye ≤∑
e∈δ+(i)
xe, ∀i ∈ Vs, (4.7)
ye ≥ 0, ∀e ∈ Ed. (4.8)
For a given z ∈ REd and d ∈ RVd , we would like to know under which conditions would there exist afeasible first-stage solution x that satisfies the system of inequalities. Note that by Farkas’ lemma,
this is the case if for all feasible values (v,w) of the dual recession cone, i.e.,
vj − vi ≤ wij , ∀(i, j) ∈ Ed, (4.9)
v,w ≥ 0 (4.10)
satisfy the following condition
∑j∈Vd
vjdj −∑i∈Vs
vi
∑e∈δ+(i)
xe
≤ ∑e∈Ed
zewe. (4.11)
Following Atamtürk and Zhang (2007), we observe that (4.9)–(4.10) is the dual recession cone of
a subgraph G(V,Ed), where V = Vs ∪ Vd. The extreme rays of the dual recession cone correspondto cuts of the subgraph. In other words, (v,w) is an extreme ray if there exists a node set S ⊆ Vwhere vj = 1 if j ∈ S, vj = 0 if j /∈ S, wij = 1 if i /∈ S, j ∈ S, and wij = 0 otherwise. Therefore,
15
-
Figure 8: Graph partitioning using k-means with 8
clusters. Nodes with the same color belong to the same
cluster. The red nodes correspond to cluster centers.
Lines connecting nodes refer to a direct connection be-
tween the two nodes.
it is sufficient to check (4.11) for all extreme rays in the dual recession cone. Hence, for a given
z ∈ REd and d ∈ RVd , x is a feasible first-stage flow if and only if∑i∈Ss
∑e∈δ+(i)
xe +∑i/∈Ss
∑j∈Sd
zij ≥∑j∈Sd
dj , ∀Ss ⊆ Vs, Sd ⊆ Vd
Hence, consider the following optimization problem with an exponential number of constraints:
minimizex,z
∑e∈Ed
ceze
subject to∑
e∈δ−(i)
xe ≤ Si, ∀i ∈ Vw,
∑i∈Ss
∑e∈δ+(i)
xe +∑i/∈Ss
∑j∈Sd
zij ≥ maxd∈U
∑j∈Sd
dj
, ∀Ss ⊆ Vs, Sd ⊆ Vd,xe ≥ 0, ∀e ∈ Es,
ze ≥ 0, ∀e ∈ Ed.
(4.12)
Note that the optimal cost of (4.12) is an upper bound for the cost of the original min-max-min
robust optimization model (4.4). Intuitively, this is because it is a more restrictive problem since
ze ≥ ye(d) for all d [Note that ye depends on the particular realization of d due to its being arecourse decision variable].
Note that this optimization model has an exponential number of constraints. For example, if
there are 10 temporary storage areas and 10 demand sites, the order of magnitude of the number
of constraints is 210×210 ≃ 106. If there are 30 temporary storage areas and 30 demand sites, thenthe number of constraints is roughly 1018!
16
-
We make the optimization model more tractable in a three-phased approximation, which we
describe next. We illustrate the phases in computational experiments with data from Typhoon
Haiyan. For interested readers, we describe in Appendix A how the model parameters are deter-
mined from raw data.
4.1.1 Phase I: Geo-location clustering of municipalities
In the first phase, we reduce the number of nodes by clustering the transportation graph by location
into k clusters. For example, Figure 8 shows a partition with 8 clusters. The partitioning can be
calculated by solving the k-center problem. That is, given n cities with specified distances, one
wants to build k warehouses in different cities and minimize the maximum distance of a city to a
warehouse. In graph theory, this means finding a set of k vertices for which the largest distance of
any point to its closest vertex in the k-set is the smallest. The graph must be a complete graph
with distances satisfying the triangle inequality. In our setting, these conditions can be fulfilled
by adding missing edges with distance equal to the shortest distance between the two endpoints.
This problem is NP-hard. However, simple greedy approximation algorithms can build a k-set with
an approximation factor of 2. See Vazirani (2001) for more information on the k-center problem.
However, another simpler method for generate reasonable partitions is through k-means clustering,
popular in data-mining, where each city is an observation represented by its longitude and latitude.
The goal of k-means clustering is to partition the n observations into sets so as to minimize the
within-cluster sum of squares. Note that the difference with the k-means clustering is that it ignores
the edge information (actual distances) and bases its clustering on the Manhattan distance. For
example, k-means will put two cities “close” to each other in the same cluster, even if the road
between them must go around a mountain. However, based on applying k-means clustering to
actual data, we believe that clustering using k-means results in good partitions (see Figure 8).
4.1.2 Phase II: Solve the inter-cluster problem
We assume that there is a single demand site and a single temporary storage site in each cluster
located at the cluster center. Instead of solving the original problem with |Vd| demand nodes and|Vs| storage nodes, we can solve the much smaller problem instance of (4.12) with k demand sitesand k temporary storage locations. The first stage solution will prescribe storage quantities for
each cluster.
Figure 9a plots the solution of the inter-cluster problem (4.12) with 8 clusters. The green nodes
correspond to locations of stockpiled relief items in regional warehouses (Cebu City, Iloilo City
and Tacloban City). The yellow nodes correspond to the 8 cluster centers, and the size of the
node is proportional to the supply to be pre-positioned in the cluster. The Iloilo City warehouse
ships family food packs for pre-positioning in three clusters in Western Visayas. The Cebu City
warehouse ships FFPs to serve four small clusters in Central and Eastern Visayas. The Tacloban
City warehouse ships FFPs to serve two clusters in Eastern Visayas.
17
-
(a) Inter-cluster allocation (b) Intra-cluster pre-positioning
Figure 9: The solutions of inter-cluster allocation and intra-cluster pre-positioning on Typhoon Haiyan data. In (a), the green
nodes correspond to supply in regional warehouses. The yellow nodes correspond to supply allocated to clusters. The dotted
line between a regional warehouse and cluster center imply that the warehouse ships supply to that cluster. In (b), the yellow
nodes correspond to pre-positioned supply in temporary storage facilities. The red nodes are municipalities with demand.
4.1.3 Phase III: Solve the intra-cluster pre-positioning problem
After Phase II, we have determined the quantities of relief items to pre-position in a given cluster.
However, we still have not determined the location of where to pre-position the relief items. To
do this, we solve an intra-cluster pre-positioning problem, which determines where and how much
relief items to pre-position in a set of candidate temporary storage facilities. Let C be the cluster,
and Vs(C) be the set of candidate temporary storage facilities. We can assume that the within the
cluster, demand is deterministic (since there’s less demand variation within a cluster). Suppose S
is the supply to be pre-positioned in a cluster (from the solution of Phase II). Given that we want
to set up at most M temporary storage facilities in the cluster, the intra-cluster pre-positioning
18
-
problem solves the deterministic problem:
minimizes,y,z
∑e∈Ed(C)
ceye
subject to∑
e∈δ+(j)
ye + sj ≥ dj +∑
e∈δ−(j)
ye, ∀j ∈ Vs(C),
∑e∈δ+(j)
ye ≥ dj , ∀j ∈ C \ Vs(C),
∑j∈Vs(C)
sj = S,
sj ≤ Szj , ∀j ∈ Vs(C),∑j∈Vs(C)
zj ≤ M,
ye ≥ 0, ∀e ∈ Ed(C),
zj ∈ {0, 1} , ∀j ∈ Vs(C).
(4.13)
The problem solves for the optimal pre-positioned supply at a municipality j, denoted by sj . Note
that zj is a binary decision variable that takes a value of 1 if temporary storage facility j is used
for pre-positioning relief items, and takes a value of 0 otherwise. The first set of constraints specify
that flow into a municipality plus pre-positioned supply should cover both demand and outflows
for municipalities that are candidate storage areas. The second set of constraints specify that non-
storage municipalities must have enough inflows to meet the demand. The third constraint says
that the sum of pre-positioned supply in all municipalities must equal S. Because of the fourth
constraint, a municipality can only have pre-positioned supply if zj = 1. The fifth constraint
specifies that there can be at most M temporary storage locations in the cluster.
A challenge in the intra-cluster problem is the which values for the demand to use in param-
eterizing the model (4.13). One solution is to simply use the nominal demands d̄. However, the
problem with this is that there might not be enough supply S planned for the cluster to meet the
nominal demand (since it was allowed to deviate from nominal in the intra-cluster phase). Instead,
one can scale the nominal demand to achieve a feasible problem (4.13). Alternatively, one can set
y = z in the recourse problem, where z is the optimal value of the variables in (4.12) (i.e., assume
that regardless of the demand realization, the recourse is the same), and determine the minimum
feasible cluster demands. Then scale the municipality demands based on the minimum feasible
cluster demands.
Solving (4.13) on Typhoon Haiyan data, with a maximum of two temporary storage facilities
per cluster, we get the solution illustrated in Figure 9a. The yellow nodes correspond to the
temporary storage facilities with pre-positioned relief items. The size of the nodes is proportional
to the pre-positioned quantity. There are a total number of 16 pre-positioning facilities utilized.
The red nodes correspond to municipalities with post-disaster demand. The size of the red nodes
is proportional to the demand. Tacloban City has the largest quantity of pre-positioned supply.
19
-
4.2 Fair and efficient distribution of FFPs post-disaster
Rapid damage and needs assessment (RDNA) is done post-disaster to determine affected areas and
populations with the greatest unmet needs. Once needs have been assessed, disaster managers have
to decide on how to allocate the available emergency relief items to the affected population. A tool
that has increasingly been employed in such assessments is a “prioritization matrix” which ranks
municipalities in priority for receiving aid (Benini and Chataigner, 2014). This matrix contains
indicators which are aggregated into a composite index that determines the rank of the municipal-
ity. Prioritization matrices were used in disaster response for Yolanda. These matrices combine
indicators for magnitude (e.g. number of affected persons), intensity (e.g. percentage of destroyed
dwellings, wind speed), and pre-existing conditions (e.g. poverty rate). Prioritization matrices are
useful visual guides for disaster managers to make post-disaster distribution decisions.
However, despite having this information, post-disaster distribution remains a difficult task es-
pecially with a large-scale disaster such as Yolanda since multiple facilities have to be coordinated
for a timely distribution. Much of the focus of post-disaster operations after Typhoon Yolanda have
been towards Tacloban City, the hardest hit area, which ignored other areas that were needing re-
lief aid. A UN assessment of nine municipalities in Leyte two weeks after the typhoon (UNOCHA
OCHA, 2013) reported that “food appears to be effectively distributed in some [areas], but not
effectively or evenly distributed in others...[T]he more remote communities are not notified ade-
quately or are required to walk in for what remains available.” As noted in an Oxfam briefing after
Yolanda (Oxfam, 2013), “a wider problem is that the levels of assistance vary depending on loca-
tion...Relief operations largely began in urban areas and then spread slowly into adjoining districts
and beyond, but the limited number of heavy vehicles and warehouse facilities in hubs such as
Tacloban has delayed distribution.”
In what follows, we develop an optimization model which determines the optimal post-disaster
distribution from supply hubs to affected municipalities and determines the optimal positioning
of transportation vehicles. The model optimizes fairness of distribution (by using an objective
that considers backlog of municipalities weighted by the composite indices from a prioritization
matrix), as well as timeliness (by coordinating facilities and appropriately stationing vehicles).
Pre-positioning family food packs (from Section 4.1) achieved the goal of reducing response time.
However, after a post-disaster distribution has been planned, the timeliness is also affected by how
the limited vehicles are stationed. Clearly, there will be shorter response times in areas where the
hubs have many stationed vehicles.
We first introduce the notation that will be used in our model. Let T be the number of time
periods in the disaster response phase. Let V be the set of municipalities. We denote by qj the
quantity of pre-positioned supply at municipality j ∈ V , where qj = 0 if there is no pre-positionedsupply. Let M be the set of transportation modes (e.g. truck, plane, ship). We denote by Um the
total number of vehicles of mode m ∈ M that have to be stationed in hubs, and Vm ⊆ V is theset of municipalities that are candidate hubs for vehicles of mode m. If a vehicle of mode m is
stationed at hub j, then we denote by µmj the average service time (in days per unit), which is the
20
-
average time the vehicle is in use in one roundtrip divided by the average load size. The number
of vehicles of mode m that are stationed in hub j is umj .
Let us focus on a given municipality j ∈ V . At the start of each period, municipality j receivesdeliveries coming from other municipalities and sends deliveries to other municipalities. We denote
by fijtm the number of units to be sent from municipality i to j at the start of period t with
transportation mode m. After the deliveries are completed, the net inventory at municipality j,
which we denote by Ijt, is used to meet new demand, which we denote by djt. If the demand is
less than the available supply, then there is a surplus at the end of period t, and we denote the
total surplus as Yjt. If the demand exceeds the available supply, then there is a backlog at the end
of period t, and we denote the total backlog as Bjt. Backlog or surplus at the end of period t is
carried over to the next period t+ 1.
The following table summarizes the notation that we discussed above:
Parameters:
V the set of municipalities
λj the weighted preferential composite index of municipality j
qj the quantity of pre-positioned supply at municipality j ∈ V
djt the new demand at municipality j at period t
M the set of transportation modes
Vm the set of candidate hubs for vehicles of transportation mode m
Um the number of transportation vehicles of mode m
µmj the average service time (days per unit) of a transportation vehicle of mode m stationed at hub j
Decision variables:
fetm the number of units sent on edge e at the start of period t using transportation mode m
umj the number of vehicles of transportation mode m stationed at hub j
Ijt the net inventory at municipality j at period t before account for new demand
Bjt the backlog at municipality j at the end of period t
Yjt the surplus at municipality j at the end of period t
Let us now develop the objective function that we will use for the post-disaster distribution
problem. An important quantity for a municipality is the current percentage backlog, which is de-
fined as the current backlog divided by the cumulative demand. Using our notation, the percentage
cumulative demand at municipality j at period t is:
Bjt∑s≤t djs
.
21
-
The percentage backlog is an indicator for whether a municipality’s needs are being met by the
distribution of relief goods. The percentage backlog can take any value between 0 and 1. If
the percentage backlog is 1, then this implies that there has been no inflow of supply into the
municipality since the start of the post-disaster period. If the percentage backlog is 0, then all of the
municipality’s needs post-disaster have been met. However, as has been discussed earlier, a disaster
manager might have different prioritizations in post-disaster response for different municipalities.
We take this into account by minimizing, for each time period t, the weighted sum of percentage
backlogs
∑j∈V
(λj ×
Bjt∑s≤t djs
),
where λj is prioritization composite index of municipality j (higher values of λj denote municipal-
ities with high prioritization). By weighing the backlog by the index, the optimization model will
try to reduce backlogs in municipalities with a higher value index value.
Hence, the post-disaster distribution optimization model solves:
minimizeu,B,Y,I,f
T∑t=1
∑j∈V
(λj ×
Bjt∑s≤t djs
),
subject to Ijt − djt = Yjt −Bjt, ∀j ∈ V, t ∈ [1, T ],
Ijt = Yj,t−1 −Bj,t−1 +∑m∈M
∑e∈δ+(j)
fetm −∑m∈M
∑e∈δ−(j)
fetm, ∀j ∈ V, t ∈ [1, T ],
Yj0 = qj , Bj0 = 0, ∀j ∈ V,∑m∈M
∑e∈δ−(j)
fetm ≤ Yj,t−1, ∀j ∈ V, t ∈ [1, T ],
∑e∈δ−(j)
fetm ≤ umj/µmj , ∀m ∈ M, j ∈ Vm,
∑j∈Vm
umj = Um, ∀m ∈ M,
umj ∈ Z+,B ≥ 0,Y ≥ 0, f ≥ 0.(4.14)
The first constraint is that in a given period and municipality, there either is a surplus or a backlog.
The second constraints are the inventory dynamics, where the net inventory available in a period
consists of backlog or surplus carried over from the previous period plus inflows minus outflows.
The third constraints are the initial conditions for surplus and backlog in a municipality. The fourth
constraints specify that the total outflows in a municipality cannot exceed the surplus inventory
carried over from the previous period. The fifth constraints specify that the total outflow from a
municipality using a particular transportation mode cannot exceed the installed capacity weighted
by the average service time. The sixth constraint specifies that the sum of installed capacity over
all municipalities must equal the available capacity.
22
-
(a) Trucks (b) Ships (c) Planes
Figure 10: The solutions of post-disaster vehicle positioning. Nodes with the same color belong in the same cluster. Red nodes
are hubs for transportation vehicles.
Optimization problem (4.14) is a tractable mixed integer linear program which can be solved
using commercial off-the-shelf solvers such as ILOG CPLEX. We ran the optimization problem on
data from Typhoon Haiyan and Figure 10 shows the transportation hubs based on the solution of
the model.
5 Computational experiments from Typhoon Yolanda data
On November 8, 2013, Typhoon Yolanda made landfall in the Visayas islands in the Philippines,
with strongest winds in recorded history of 195 mph. In the typhoon’s aftermath, there have been
an estimate of more than 6,000 casualties and more than 14 million people affected. In this section,
we will further discuss details of the computational experiments of the optimization models on data
gathered from Typhoon Yolanda in 2013.
5.1 Data Gathering
Data gathering was a long process that took a few weeks due to the lack of available data, and
that we had to assemble the data from several different sources. In what follows, we describe to
the best extent possible the steps we took to gather data that we use for the experiments.
5.1.1 Municipality properties
We gathered information on municipalities that were affected by Typhoon Haiyan from several
public sources (NSCB, 2014; PhilGIS, 2013). The following table is representative of the type of
information that we gathered:
23
-
Code Municipality Province Region Longitude Latitude Income Class Population
83701000 Abuyog Leyte VIII 125.03 10.68 - 57,146
63001000 Ajuy Iloilo VI 123.02 11.15 2nd 47,248
83702000 Alangalang Leyte VIII 124.86 11.21 2nd 46,411
83703000 Albuera Leyte VIII 124.73 10.91 - 40,553
71201000 Albuquerque Bohol VII 123.96 9.63 5th 9,921
We used the NSCB municipality code as a unique identifier for a municipality for the computational
experiments.
The next data we collected was to recreate the Visayas transportation network (road, sea routes,
and air routes). Given the lack of data on the road network in Visayas, we used Google Maps to
query information on national highway road distances and connections between municipalities, and
enter them manually on to a csv file. It only took a few days to complete this procedure since
the major road network in Visayas is very sparse. For future implementations, it is possible to
automize this step by creating a script that automatically scrapes data from Google Maps. The
following table shows examples of the information that we gathered:
Municipality 1 Province 1 Code 1 Municipality 2 Province 2 Code 2 Road distance (km)
Buruanga Aklan 60405000 Malay Aklan 60412000 6.5
Malay Aklan 60412000 Nabas Aklan 60414000 26.2
Nabas Aklan 60414000 Ibajay Aklan 60406000 11.1
Ibajay Aklan 60406000 Tangalan Aklan 60417000 15.2
Nabas Aklan 60414000 Pandan Antique 60611000 11.5
We used Ports.com to find information on seaports and existing sea routes and their distances
in the Visayas region. There are a total of 46 municipalities with seaports. We used the distance
information to create a matrix of sea route distances. The following is part of the sea route distance
(in miles) matrix:
Seaport Allen Bacolod Bais Barcelona Baybay Borongan
Seaport 84801000 64501000 74604000 64509000 83708000 82604000
Allen 84801000 - - - - - -
Bacolod 64501000 - - - 44.82 - -
Bais 74604000 - - - 95.63 - -
Barcelona 64509000 - 44.82 95.63 - - 139.24
Baybay 83708000 - - - - - -
Borongan 82604000 - - - 139.24 - -
To gather data on air routes, we used the Visayas Logistics Map (see Figure 3b) to list the
municipalities that the map shows have airports or airstrips. There are a total of 23 municipalities
with airports or airstrips. For simplicity, we assume that there is an air route between any two
municipalities in this list, and the distance of the air route is simply the straight line distance.
Figure 11 shows a visual representation of the data we gathered on the Visayas transportation
network. The grey nodes are the municipalities plotted based on their geographical coordinates on
24
-
Figure 11: Grey nodes represent municipalities. Blue nodes
represent municipalities with a seaport. Green nodes rep-
resent municipalities with an airport. Red nodes represent
municipalities with a seaport and an airport. Grey edges
connecting municipalities represent major roads.
the map. The blue nodes represent municipalities containing seaports. The green nodes represent
municipalities containing an airport or airstrip. The red nodes are municipalities with both a
seaport and an airport/airstrip. Grey lines connecting municipalities are segments of major roads
or highways.
5.1.2 Typhoon Haiyan supply and demand
We compiled data on the total affected population by municipality from Typhoon Haiyan based on
Situational Reports released by NDRRMC (NDRRMC, 2013). All NDRRMC Situation Reports
can be accessed through ReliefWeb.int. A total of 44 provinces were affected by the typhoon, 574
municipalities in Regions IV-A, IV-B, V, VI, VII, VIII, X, XI and CARAGA. The data we gathered
is in the following format:
Municipality Province Code Affected Families Affected Persons
Altavas Aklan 60401000 5,450 26,905
Balete Aklan 60402000 5,409 27,045
Banga Aklan 60403000 8,281 33,955
Batan Aklan 60404000 8,171 34,084
Buruanga Aklan 60405000 3,000 17,213
We assume that there is one regional warehouse per affected region by Typhoon Yolanda.
5.2 Methodology
We now discuss the methodology in the computational experiments. For simplicity, we isolate the
computational experiments to the municipalities that are within 100 kilometers from the path of
the typhoon.
25
-
5.2.1 Pre-positioning FFPs in temporary storage
• We implemented the three-phase approximation to the pre-positioning problem in Section 4.1with the data. First, we define the clusters for each municipalities using k-means with 8
clusters. Figure 8 shows the result of the clustering algorithm. The details of the clusters in
the figure are given in the following table:
Cluster Color Cluster Center No. of Municipalities
Salmon Alimodian, Iloilo 26
Deep sky blue Jaro, Leyte 30
Pale green Tuburan, Cebu 34
Khaki Daram, Western Samar 38
Slate blue Malinao, Aklan 22
Plum Lemery, Iloilo 39
Tan Sogod, Southern Leyte 29
Silver Gen. MacArthur, Eastern Samar 17
• We created a master transportation graph with arcs corresponding to road connections, seaconnections, and air connections. The distances in each arc is in travel time (assuming trucks
travel in roads, ships travel in sea, and planes travel by air).
• We created a graph for the inter-cluster pre-positioning problem with one warehouse node foreach regional warehouse, one temporary storage node in each cluster center, and one demand
site in each cluster center.
• First stage arcs between a regional warehouse to a temporary storage nodes are created ifthere exists a transportation path between the two (through a combination of road, sea route,
or air route). The arc distance between the two is equal to the length of the shortest path in
the transportation graph. A second stage arc is created for each combination of temporary
storage location and demand site for which there is a path. The distance of the arc is equal
to the length of the shortest path in the transportation graph.
• The nominal demand in each demand site is equal to the cluster’s aggregate nominal demand.We assume a demand uncertainty set that allows demand with no more than a 20% average
deviation from the nominal demand.
• We assume that each regional warehouse stores 5 million units each.
• We solve the inter-cluster pre-positioning problem (4.12). The solution is visualized in Fig-ure 9a. The details of the solution is shown in the following table:
26
-
Cluster Color Cluster Center Units from Tacloban Units from Cebu Units from Iloilo
Salmon Alimodian, Iloilo 0 0 2,804,427
Deep sky blue Jaro, Leyte 0 1,110,163 0
Pale green Tuburan, Cebu 0 0 635,824
Khaki Daram, Western Samar 0 322,547 0
Slate blue Malinao, Aklan 0 324,425 0
Plum Lemery, Iloilo 4,767,081 0 0
Tan Sogod, Southern Leyte 0 0 1,451,743
Silver Gen. MacArthur, Eastern Samar 293,918 62,258 0
• We solve the intra-cluster pre-positioning problem (4.13) given the cluster supplies in thetable above. Assuming that each cluster can only have a maximum of two storage areas.
The candidate storage locations that we chose are municipalities that have a National Food
Authority (NFA) warehouse. The total number of NFA warehouses in the regions affected
by the typhoon are 44. Figure 9b visualizes the solution of the intra-cluster pre-positioning
problem.
5.2.2 Distribution of FFPs post-disaster
• We assume that there are five days in the post-disaster distribution time horizon.
• We assume that new demand occurs every other day, since family food packs are assumed toaddress the needs of a family of five for two days.
• We assume that the pre-positioned supply of the previous computational experiments willbe available for the system post-disaster. We will call locations with pre-positioned supply a
“supply hub”.
• We assume three types of transportation vehicles: trucks, ships, planes. We will call locationsthat are airports, air strips, or seaports as “transportation hubs”. The following table shows
the values for the load size, time to load, and the total number of vehicles:
Vehicle Mode Speed Load size Loading time Number of vehicles
Truck Land 80 kph 3,000 0.5 hr 50
Ship Sea 20 kph 160,000 10.0 hr 5
Plane Air 480 kph 2,000 0.5 hr 5
• We created an directed graph where there is a node for every municipality. Edges originatefrom a supply hub or a transportation hub to any municipality that has a path to it. The
length of the edge is equal to the shortest path distance in the transportation graph.
• Any supply hub or transportation hub is a candidate location for a vehicle. We calculate theaverage service times from a supply hub or a transportation hub j using transportation mode
27
-
m by:
µmj =2×Ave. distance of j to neighbors + Loading time of m
Load size of m
• Using these model parameters, we solve (4.14). The model’s recommended location of trans-portation hubs is visualized in Figure 10. The distribution plan is too complicated to visualize.
6 Decision Support Tool Implementation
The optimization models developed in Section 4 can be used as a building block for a decision sup-
port tool for disaster managers in pre-disaster and post-disaster operations. Figure 12a illustrates
one simple embodiment of a decision support tool, which we implemented with a user interface in
Excel. The user can input model parameters, including model tolerances, the prioritization indices
for municipalities, the list of candidate storage ares, the list of candidate transportation hubs, and
vehicle speeds and capacities. The model also reads csv files containing details of municipality
properties, transportation arc distances, and demand and arc capacity forecasts. The inputs are
read by a computer program implemented in Python that uses the input to estimate the model pa-
rameters. Then the program creates a CPLEX optimization model with these parameters through
the Python CPLEX API. The model takes only a few seconds to solve and the output is plotted
and logged into a csv file.
In a project funded by the Department of Science and Technology entitled “Pre-positioning and
allocation of relief supplies GIS-based support system for disaster preparedness and response” (a
project in collaboration with Ateneo de Manila University and IBM Philippines), we plan for the
optimization model to be one of the fundamental components in a cloud-based decision support
tool for disaster planning, which couples numerical weather models, machine learning models for
prediction, and the optimization models. Figure 12b illustrates the envisioned system context
diagram. In this system, the tool will automatically fetch data from DOST and DSWD IT systems.
However, in case live data is unavailable, the tool can still function independently by allowing users
to encode data manually, import from files, or even use previous data stored in the cloud. Reports
and results of simulation can be exported and viewed locally without internet connection using
document viewers and map viewers like Google Earth. All sensitive data will be encrypted in the
database and user access is restricted through authentication.
7 Conclusions
Efficient humanitarian operations is a pressing need in the disaster-prone Philippines. In this
report, we introduced optimization models for efficient pre-positioning, supply allocation and vehicle
positioning. There are many other challenges in analytics for humanitarian operations which have
been touched upon, but not directly addressed in this report. Firstly, demand forecasting is difficult,
but crucial for planning. There are sophisticated weather forecasting models available to predict
28
-
(a) Simple embodiment (b) Cloud-based embodiment
Figure 12: System context diagrams
typhoon tracks, however these forecasts are subject to high uncertainty, and the weather forecasts
have to be converted to demand forecasts to be used in supply chain models. In this report,
we addressed demand uncertainties by using ideas from robust optimization by creating demand
uncertainty sets. Another challenge is that supply is uncertain, due to supply being not only from
prepositioned supply, but also from irregular donations. A last challenge is that in the humanitarian
logistics landscape, there are multiple humanitarian organizations (NGOs) operating in parallel in
the same landscape. Hence, there is a risk of duplicating efforts.
29
-
References
A. Apte. Humanitarian logistics: A new field of research and action. Foundations and Trends in Technology,
Information and Operations Management, 3(1):1–100, 2009.
A. Atamtürk and M. Zhang. Two-stage robust network flow and design under demand uncertainty. Operations
Research, 55(4):662–673, 2007.
B. Balcik and B. M. Beamon. Facility location in humanitarian relief. International Journal of Logistics:
Research and Applications, 11(2):101–121, 2008.
B. M. Beamon and S. A. Kotleba. Inventory management support systems for emergency humanitarian relief
operations in South Sudan. International Journal of Logistics Management, 17:187–212, 2006.
A. Benini and P. Chataigner. Composite measures of local disaster impact - Lessons learned from Ty-
phoon Yolanda, Philippines. http://www.acaps.org/img/documents/c-140527_compositemeasures_
philippines.pdf, May 2014. Accessed on 2014-12-29.
M. Bueza. The role of LGUs, local councils during disasters. http://www.rappler.com/newsbreak/44026-
role-lgu-local-councils-disasters, November 2013. Accessed on 2014-02-19.
R. K. Cheung and C. Y. Chen. A two-stage stochastic network model and solution methods for the dynamic
empty container allocation problem. Transportation Science, 32(2):142–162, 1998.
DSWD. Dswd clarifies report of expired relief noodles in negros occidental. http://www.dswd.gov.ph/
2014/12/dswd-clarifies-report-of-expired-relief-noodles-in-negros-occidental/, December
2014. Accessed on 2014-12-29.
R. Francisco. Philippines start evacuating residents ahead of typhoon Hagupit. Reuters. http://www.
reuters.com/article/2014/12/03/us-philippines-typhoon-hagupit-idUSKCN0JH0NG20141203, De-
cember 2014. Accessed on 2014-12-29.
A. Fritz. Landfall or not? Forecasts disagree on path of dangerous Typhoon Hagupit as it nears the
Philippines. Washington Post Blog. http://www.washingtonpost.com/blogs/capital-weather-
gang/wp/2014/12/03/landfall-or-not-forecasts-disagree-on-path-of-dangerous-typhoon-
hagupit-as-it-nears-the-philippines/, December 2014. Accessed on 2014-12-29.
NDRRMC. Ndrrmc update sitrep no. 29 effects of typhoon “yolanda” (haiyan). http://reliefweb.int/
report/philippines/ndrrmc-update-sitrep-no-29-effects-typhoon-yolanda-haiyan, November
2013. Accessed on 2014-12-29.
NSCB. Philippine statistics authority - national statistical coordination board: Active stats. http://www.
nscb.gov.ph/activestats/psgc/listreg.asp, September 2014. Accessed on 2014-12-29.
Oxfam. Typhoon Haiyan: the response so far and vital lessons for the philippines re-
covery. http://www.oxfam.org/sites/www.oxfam.org/files/bn-typhoon-haiyan-philippines-
response-071213-en.pdf, December 2013. Accessed on 2014-12-29.
PhilGIS. Philippine GIS data clearinghouse. Country Towns and Cities, http://www.philgis.org/
freegisdata.htm, December 2013. Accessed on 2014-12-29.
30
http://www.acaps.org/img/documents/c-140527_compositemeasures_philippines.pdfhttp://www.acaps.org/img/documents/c-140527_compositemeasures_philippines.pdfhttp://www.rappler.com/newsbreak/44026-role-lgu-local-councils-disastershttp://www.rappler.com/newsbreak/44026-role-lgu-local-councils-disastershttp://www.dswd.gov.ph/2014/12/dswd-clarifies-report-of-expired-relief-noodles-in-negros-occidental/http://www.dswd.gov.ph/2014/12/dswd-clarifies-report-of-expired-relief-noodles-in-negros-occidental/http://www.reuters.com/article/2014/12/03/us-philippines-typhoon-hagupit-idUSKCN0JH0NG20141203http://www.reuters.com/article/2014/12/03/us-philippines-typhoon-hagupit-idUSKCN0JH0NG20141203http://www.washingtonpost.com/blogs/capital-weather-gang/wp/2014/12/03/landfall-or-not-forecasts-disagree-on-path-of-dangerous-typhoon-hagupit-as-it-nears-the-philippines/http://www.washingtonpost.com/blogs/capital-weather-gang/wp/2014/12/03/landfall-or-not-forecasts-disagree-on-path-of-dangerous-typhoon-hagupit-as-it-nears-the-philippines/http://www.washingtonpost.com/blogs/capital-weather-gang/wp/2014/12/03/landfall-or-not-forecasts-disagree-on-path-of-dangerous-typhoon-hagupit-as-it-nears-the-philippines/http://reliefweb.int/report/philippines/ndrrmc-update-sitrep-no-29-effects-typhoon-yolanda-haiyanhttp://reliefweb.int/report/philippines/ndrrmc-update-sitrep-no-29-effects-typhoon-yolanda-haiyanhttp://www.nscb.gov.ph/activestats/psgc/listreg.asphttp://www.nscb.gov.ph/activestats/psgc/listreg.asphttp://www.oxfam.org/sites/www.oxfam.org/files/bn-typhoon-haiyan-philippines-response-071213-en.pdfhttp://www.oxfam.org/sites/www.oxfam.org/files/bn-typhoon-haiyan-philippines-response-071213-en.pdfhttp://www.philgis.org/freegisdata.htmhttp://www.philgis.org/freegisdata.htm
-
W. B. Powell and R. K. Cheung. A network recourse decomposition method for dynamic networks with
random arc capacities. Networks, 24:369–384, 1994.
A. Thomas and M. Mizushima. Logistics training: necessity or luxury? Forced Migration Review, 22(1):
60–61, 2011.
UNOCHA OCHA. Typhoon Haiyan [Yolanda] rapid needs assessment report: 9 municipalities in
Leyte province 25-30 November 2013. http://mhpss.net/wp-content/uploads/group-documents/219/
1386012339-RapidAssessmentReport_NineMunicipalitiesinLeyteProvince.pdf, November 2013.
Accessed on 2014-12-29.
V. V. Vazirani. Approximation Algorithms. Springer, Berlin, 2001.
31
http://mhpss.net/wp-content/uploads/group-documents/219/1386012339-RapidAssessmentReport_NineMunicipalitiesinLeyteProvince.pdfhttp://mhpss.net/wp-content/uploads/group-documents/219/1386012339-RapidAssessmentReport_NineMunicipalitiesinLeyteProvince.pdf
Review of activities undertakenIntroduction humanitarian logisticsHumanitarian logistics in the PhilippinesThe mathematical modelsPre-positioning FFPs in temporary storage pre-disasterPhase I: Geo-location clustering of municipalitiesPhase II: Solve the inter-cluster problemPhase III: Solve the intra-cluster pre-positioning problem
Fair and efficient distribution of FFPs post-disaster
Computational experiments from Typhoon Yolanda dataData GatheringMunicipality propertiesTyphoon Haiyan supply and demand
MethodologyPre-positioning FFPs in temporary storageDistribution of FFPs post-disaster
Decision Support Tool ImplementationConclusions